A high-throughput and memory-efficient inference and serving engine for LLMs
[Model][Dflash] Enable Dflash support for Qwen3NextForCausalLM targets (#45319)
Signed-off-by: Jonas I. Liechti <j-i-l@t4d.ch>
Jonas I. Liechti committed
Jun 12, 2026 at 19:30 UTC
d6fd7ce8daccb290e10c03cbf017d1eb65be4487
2 files changed
+10
tests/models/registry.py
+8
@@ -1425,6 +1425,14 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
1425
max_model_len=8192, # Reduce max len to ensure test runs in low-VRAM CI env
1426
max_num_seqs=32,
1427
),
1428
+ "DFlashQwen3NextDraftModel": _HfExamplesInfo(
1429
+ "Qwen/Qwen3-Coder-Next",
1430
+ speculative_model="z-lab/Qwen3-Coder-Next-DFlash",
1431
+ use_original_num_layers=True, # DFlash requires all layers
1432
+ max_model_len=8192, # Reduce for CI
1433
+ max_num_seqs=32,
1434
+ min_transformers_version="4.56.3", # Required for Qwen3Next
1435
+ ),
1436
# [Eagle]
1437
"EagleCohereForCausalLM": _HfExamplesInfo(
1438
"/host/engines/cohere-moe",
vllm/model_executor/models/qwen3_next.py
+2
@@ -68,6 +68,7 @@ from .interfaces import (
68
HasInnerState,
69
IsHybrid,
70
MixtureOfExperts,
71
+ SupportsEagle3,
72
SupportsLoRA,
73
SupportsPP,
74
)
@@ -758,6 +759,7 @@ class Qwen3NextForCausalLM(
759
SupportsPP,
760
QwenNextMixtureOfExperts,
761
IsHybrid,
762
+ SupportsEagle3,
763
):
764
packed_modules_mapping = {
765
"qkv_proj": [